by Yibin in Summer 2019
Several DL models that can classify human activities from Xsens sensor data. Models are based on Residual Neural Network adpated from Resnet. It achieves 95% accuracy in the training/testing data and 90% accuracy in the task data prediction.
Pytorch, torchvision, and numpy. Recommend to install anaconda first.
- Models folder: this folder contains all the model I built. See Model Description.txt for details about each model.
- Activity Recognition from Single Chest-Mounted Accelerometer folder: this folder contains the Activity Recognition from Single Chest-Mounted Accelerometer data from UCI. Only used this dataset as a proof-of-concept to my model.
- Spinetrack Data folder: all data that download from box folder. This folder has two sub-folder: data and task. data is used for model training, and the task folder is used for task prediction. See Data Description Section for more details.
- Files ending in .ipynb: these are the jupyter notebook that have all the code. See Jupyter Notebook Section for more details.
- force regression folder: not related to this project
The whole Spinetrack Data folder is downloadable from box. data sub-folder includes all training data (calibration), while task sub-folder contains all actual task data.
each subject's calibration data is in this folder. For each person (except Niral), they all have three different data - raw data (directly under each person's folder), processed data (folder Processed), and further processed data (folder Processed-redone).
The .csv- files which are directly under the data folder are the raw .csv-files, meaning the non-processed output from the Matlab-app SpineTrackDataAcqv15.mlapp. In the processed as well as in the processed_redone folder the pauses in between the repetitions were removed using the data_process_liftingv4.m code. The difference between those folders is, that in the processed_redone folder also the starting portion of each lift was removed, whereas in processed strictly only the pauses were removed. We had problems with the deep learning model when part of the pauses were still in the data, that's why we created processed_redone.
each person's actual activities data is in this folder. Unlike the data folder that repeates the same action in one .csv file, all activities in this folder are a sequence of activities that appears in the data folder. The model only use everything.csv as the prediction criteria.
Each jupyter notebook will automatically load the data and run either training the model or predicting the result.
window_size: number of rows that combined together to form a window.
channel: default to 1, similar to channel in an image (RGB)
input: similar to an image (channel, width, height) input, the input of neural net has a tensor of size (channel, window_size, 19)
- Spinetrack Data Testing: test the model with each person's everything.csv file. everything.csv contains a sequence of different task, and timeStamps_Everything.txt has a manual label of the activities. The model uses everything.csv as a task prediction criteria.
- Spinetrack Data Testing - Loop Fast!: similar to testing except it uses a loop to find the best accuract testing window.
- Spinetrack Data Prediction: predict the activities from everything.csv.
- Spinetrack Data Training: main training file. Trained from Spinetrack Data folder.
- Spinetrack Sub-activities xxx Data Training: train model from each sub-categorry activities. For example, in the main training, carry_10 and carry_15 is the same activity as "carrying", but here depends on the file name it is reponsoble to train only one kind of activities.
- Others: not very related to this project or outdated.
If you have any question, feel free to contact Yibin at liyibin516@berkeley.edu.